README.md

---
license: mit
---

![image/png](https://cdn-uploads.huggingface.co/production/uploads/66f5800df74ecff7c0b9b64f/vgvSVmRawl-ThPrzh55Ls.png)
# Model Card: Agora-4B

## Model Summary

Agora-4B is a 4-billion parameter, BF16-precision transformer language model, designed for ethical, inclusive, and adaptive dialogue in multi-user domestic environments. Inspired by the research paper ["Plural Voices, Single Agent: Towards Inclusive AI in Multi-User Domestic Spaces"](https://doi.org/10.48550/arXiv.2510.19008), Agora-4B incorporates principles of fairness, value alignment, and accessibility to better serve diverse household users—including children, elderly, and Neurodivergent individuals.

**Repository:** [JoydeepC/Agora-4B](https://huggingface.co/JoydeepC/Agora-4B)
**Paper:** [Plural Voices, Single Agent](https://doi.org/10.48550/arXiv.2510.19008)
**Model size:** 4B parameters
**Tensor type:** BF16
**Files:** Safetensors format (2 shards, ~8.07 GB), tokenizer files, configs, chat templates, etc.

---

## Intended Use

Agora-4B is intended for use as a core assistant agent in domestic AI deployments, especially in settings with multiple users and overlapping accessibility needs. Typical scenarios include:
Domestic voice assistants which must mediate between adult, child, and elderly users
Applications where context-sensitive safety, fairness, or ethical intervention is required
Research or development in inclusive, privacy-first AI for multi-agent, multi-user environments

---

## Model Architecture & Training

**Architecture:** 4B-parameter transformer, trained with curriculum blending human and synthetic dialogue
**Objective:** Optimized for fairness, multi-value alignment, ethical compliance, and accessibility-aware conversation
**Training Data:** Curated public datasets covering mental health, eldercare, education, and moral reasoning. Enhanced with fairness-aware, multi-user scenarios and privacy-centric synthetic examples.
**Ethical Safeguards:** Includes adaptive safety scaffolds (e.g., age-specific explanations, guidance for Neurodivergent users), autonomy sliders, and safe conflict resolution.

---

## Key Features

**Real-Time Value Alignment:** Dynamically identifies and negotiates conflicting user needs, values, and accessibility requirements
**Inclusive Design:** Special handling for overlooked populations (children, elderly, Neurodivergent), including step-by-step instructions, accessible language, and equitable interaction
**Privacy-Focused:** Avoids unnecessary data retention or sharing
**Adaptivity:** Safety, autonomy, and guidance dynamically adjusted per user/context
**Design Innovations:** Video guidance, autonomy sliders, family hubs, adaptive dashboards
**Performance:** Outperforms baselines in compliance, fairness, and safety (see paper for details)
  - Compliance: 76% (vs 70% baseline)
  - Fairness: 90% (vs 85% baseline)
  - Safety violations: 0% (vs 7% baseline)

---

## Citation

If you use this model, please cite:

```bibtex
@misc{chandra2025pluralvoicessingleagent,
      title={Plural Voices, Single Agent: Towards Inclusive AI in Multi-User Domestic Spaces}, 
      author={Joydeep Chandra and Satyam Kumar Navneet},
      year={2025},
      eprint={2510.19008},
      archivePrefix={arXiv},
      primaryClass={cs.HC},
      url={https://arxiv.org/abs/2510.19008}, 
}
```
---

## Further Reading

[arXiv:2510.19008](https://arxiv.org/abs/2510.19008)
[Project repository (HuggingFace)](https://huggingface.co/JoydeepC/Agora-4B)

---

This model and codebase are open sourced for reproducibility and collaborative research on inclusive, agentic AI.